osmosMind
← Leaderboard

health_bench

Ranked by Score·7 models·Updated June 2026

#ModelScore
1
Anthropic

Claude Opus 4.8

Anthropic

71.2
2
OpenAI

GPT-5.4

OpenAI

69.8
3
Google

Gemini 3 Pro

Google DeepMind

67.1
4
Anthropic

Claude Sonnet 4.6

Anthropic

64.5
5
DeepSeek

DeepSeek V4

DeepSeek

60.3
6
AlibabaCloud

Qwen3 Max

Alibaba

57.8
7
Meta

Llama 4.1 405B

Meta

53.2

Scores are illustrative sample data.

What it measures

health_bench evaluates health and clinical reasoning against rubrics written by practicing physicians, with explicit compliance and safety boundaries baked into the scoring.

Methodology

  • Physician rubricsPracticing clinicians set, case by case, what a correct, safe, appropriately-hedged answer looks like.

  • Safety boundariesCross a compliance or safety line and the score drops, however factually right the rest may be.

  • Dual reviewDisputed cases get a second independent pass, with inter-rater agreement reported beside the score.


← Back to Leaderboard